Release Date: Unreleased
Development Changelog: dev
This release rebuilds the package around the IsoformicExperiment class.
Table-based helpers remain available as data-frame methods, but every
analysis and plot is now first a method on the object.
IsoformicExperiment is an S7 class modelled after SummarizedExperiment
with three consistent levels: transcript assays, gene assays (derived by
summing isoforms, cached, overridable with gene_assays<-) and exon
structures from the annotation. Properties are validated on every change.prepare_isoformic_annotation(), with n_exons, tx_length,
metadata and sequence region tables). The parser is vocabulary-agnostic:
besides GENCODE GFF3/GTF it reads Ensembl and Ensembl Rapid Release GFF3,
NCBI RefSeq GFF3 and the T2T-CHM13 Liftoff/CAT annotations, keeping the
source_gene/source_transcript attributes of lifted annotations as
source_gene_id/source_transcript_id columns.download_reference() covers Ensembl (any release, species and
division, including Ensembl Plants/Metazoa/Fungi/Protists), any NCBI
assembly by accession, T2T-CHM13v2.0 (reference = "t2t") and MANE
releases; download_reference_url() caches any file by URL and
download_hprc_annotation() fetches the HPRC per-haplotype CAT
annotations.tags and an is_canonical flag
(MANE Select, Ensembl canonical, RefSeq Select), used by
dominant_isoform() to report dominant_is_canonical.prepare_isoformic_annotation(haplotype = )
labels one haplotype, merge_annotations() combines per-haplotype files
(HPRC CAT, trio-binned assemblies) into one annotation directory with
unique identifiers, and haplotype_summary() reports per-gene haplotype
fractions and dominant-isoform switches between haplotypes.download_reference(reference = "fantom")
fetches a FANTOM CAT level (GTF), download_fantom_cat_tables() the
companion tables, and the parser keeps geneClass (as gene and fallback
transcript type), geneCategory, DHS_type, coding_status,
cumulative_support and TIEScore; add_fantom_cat_tables() attaches
CAGE cluster promoters and the gene information table.CDS features
(cds_start, cds_end, cds_length) and keeps transcript_support_level;
transcript_features() computes ORF and UTR lengths, the 50-nt NMD rule
(stop_to_last_junction, nmd_predicted), transcription start/end groups
(alternative promoters and 3' ends), the canonical isoform of each gene
and splicing events relative to it (splicing_events(): skipped and
novel exons, retained introns, alternative splice sites, first and last
exons).transcript_productivity()
(productive, altered_productive, unproductive, noncoding,
uncertain, with the reason for each call), productive_fraction() per
sample or group with the between-group change, run_productivity_enrichment()
(gene sets ranked by change in productive fraction), poison_exons()
and highlight_poison_exons in plot_genomic_context().add_annotation_columns() attaches external per-transcript or per-gene
evidence (Ribo-seq, NMD inhibition, GTEx dominant isoforms, domain
retention, ...) to the annotation; a productivity_evidence column
overrides the computed productivity.extract_transcript_sequences() stores the spliced
sequence of every transcript in the annotation directory
(annot_data_sequences.parquet), from a genome FASTA (streamed in chunks,
gzip or plain, chromosome names matched with or without chr) or a
transcript FASTA; has_sequences() and transcript_sequences() read it.transcript_orfs() projects the canonical start
codon of each gene onto every isoform and translates it (the ORFanage
approach), so novel transcripts without CDS get an ORF; annotated CDS are
re-read from the sequence (annotation_cds_agrees), with stop codon,
genomic coordinates, uORFs, frame relative to the canonical ORF and an
NMD class (ptc, ptc_start_proximal, non_stop, long_3utr, none).
coding_potential() scores the longest ORF with the Fickett statistic,
an in-frame hexamer score and a frozen logistic model trained on GENCODE
49 (build_coding_model(), data-raw/build_coding_model.R; held-out
sensitivity 0.96, specificity 0.93).transcript_productivity() uses the sequence-derived ORFs when present
(use_sequence), reports source (sequence, annotation, evidence),
distinguishes alternative starts, in-frame changes, frameshifts and stops
in novel sequence among altered_productive isoforms, calls non-stop
decay and start-proximal stops, and applies evidence columns in a fixed
priority (productivity_evidence, validated NMD exon, Ribo-seq main ORF,
APPRIS principal).frozen_datasets(), fetch_frozen_data(),
read_frozen_data(), build_frozen_*()): APPRIS principal isoforms and
method scores plus SPADE Pfam domain coordinates (e116v50, Ensembl 116 / GENCODE 50), GENCODE
Ribo-seq ORFs phase 1 and 2, NMD-inducing exons validated by NMD
inhibition (Karousis et al. 2021), the PolyASite 2.0 atlas and GTEx v8
median transcript TPM per tissue. Files are pinned by version and MD5,
downloaded from the first mirror that has them (frozen_mirrors(): user
option, the frozen-data GitHub release, the Zenodo snapshot of the
processed files) or rebuilt from the sources; bundle_frozen_data()
assembles the Zenodo deposit and data-raw/README.md documents the
update and deposit procedure.data-raw/knockdown/isoform_features.R and ship in
inst/extdata/knockdown/stability/.read_treeterminus() reads the transcript trees of TreeTerminus (Newick;
leaves are positions in quant.sf, named from a Salmon directory) and
cut_transcript_trees() cuts them where the inferential relative variance of
the summed replicates falls below the cutoff: an isoform the reads determine
stays itself, one they do not is replaced by its lowest well-determined
ancestor, and nested choices are merged so that the groups never overlap.
The groups go to add_uncertainty(groups = ), so the confidence classes and
collapse_groups work at the resolution the data support, isoform by
isoform. Tested on real TreeTerminus output.run_terminus() runs Terminus from R (terminus group on every Salmon
directory, then the consensus terminus collapse) through condathis,
checks first that every directory has replicates and equivalence-class
weights, deletes the multi-gigabyte delta.log of each sample as soon as it
is grouped, and returns the consensus groups. On the six GSE129540 samples it
reproduces the consensus of the command-line run.read_terminus_groups() reads the transcript groups of Terminus
(terminus collapse, clusters.txt), and add_uncertainty(groups = )
accepts them (or any table of transcript_id and group_id) in place of the
groups built from the correlation of the replicates; the confidence classes,
collapse_groups and uncertainty_methods() follow the new groups. The
parser is tested on real output of Terminus 0.1.0 on Salmon 2.8.0.test_usage() tests differential transcript usage with edgeR
diffSplice, satuRn or DEXSeq (all in Suggests) on counts divided by the
overdispersion of the replicates, and reports transcript and gene level
p-values in one layout. With more than two groups it tests every level
against the first, every pair (contrasts = "pairwise") or a list of
c(numerator, denominator) pairs, and stacks the results with a contrast
column; diffSplice and satuRn fit all groups once, DEXSeq fits each
pair. contrasts = "omnibus" tests for a difference among any of the groups:
the DEXSeq likelihood-ratio test on all samples, or the Simes combination of
the contrasts against the reference for diffSplice and satuRn.read_oarfish_sc() reads the single-cell output of Oarfish
(--single-cell: .count.mtx, .barcodes.txt, .features.txt) and sums
the cells of each sample, given a table of barcodes and samples, into an
IsoformicExperiment (pseudobulk; TPM from the annotated transcript lengths).
read_oarfish_sc_counts() returns the cell-level counts as a long table.
The fixture is real Oarfish 0.10.3 output on simulated barcoded reads.read_oarfish() reads long-read quantifications from Oarfish (counts,
lengths, TPM computed from them). read_inferential_replicates(),
equivalence_support() (from the unique and ambiguous read counts),
salmon_run_qc() and add_uncertainty() accept Oarfish output prefixes,
including the Parquet bootstrap replicates (.infreps.pq). The bundled
fixture is real Oarfish 0.10.3 output on simulated nanopore reads. Mixing
Oarfish and Salmon in one call is an error (isoformic_mixed_quantifiers).inferential_overdispersion() and scale_counts() put the quantification
uncertainty into the counts as edgeR::catchSalmon() does (Baldoni et al.
2024), from the replicates stored in the object; the estimates match
catchSalmon() on a full 517,038-transcript quantification. A read-split
experiment on GSE129540 (data-raw/knockdown/validate_uncertainty.R) backs
the confidence classes: of the isoforms that are large and stable over the
draws in null contrasts, almost none are classed high, and the isoforms
called in one half of the reads reproduce in the other half less often as the
class or the inferential variance worsens. The small result tables are
bundled and shown in the uncertainty article.extract_splice_sites()
(and extract_transcript_sequences() when given a genome FASTA) store the
donor and acceptor windows of every intron in the annotation directory
(annot_data_splice_sites.parquet); splice_site_strength() scores them
with position weight matrices trained on the introns of the canonical
transcripts of the same annotation and reports weakest_donor and
weakest_acceptor per transcript, or the scores per intron. On GENCODE 50
(698,612 introns) the extraction takes about 1.5 minutes and the scoring
25 s. The scores describe the sites; on their own they separate NMD from
protein coding transcripts only slightly (AUC 0.55), so they are evidence
to read next to transcript_productivity(), not a classifier.isoform_ambiguity() has collapse_gain, the score Terminus
groups on (inferential relative variance of the sum of two isoforms minus
the mean of their own; negative when the sum is better determined).
tx_uncertainty() has low_information, TRUE for an uncertain isoform
without a single fragment of its own (the selection of mmcollapse), when
equivalence classes were read. add_uncertainty(cor_cutoff = "adaptive")
derives the correlation cutoff from the data as mmcollapse does; the
default stays -0.5, and on the six GSE129540 samples the adaptive value is
-0.25, which calls 36,910 isoforms ambiguous in place of 8,076.equivalence_identifiability() reports which transcripts the equivalence
classes of Salmon can tell apart: a transcript is identifiable when its
abundance is fixed once the total of every class is fixed. It is a
property of the observed classes, not of the posterior draws, and close to
independent of them. On the six GSE129540 samples 13,790 of 116,389
expressed isoforms are non-identifiable in every sample, and their
estimates differ more between two halves of the reads (median absolute
log2 ratio 2.4 against 0.7 at 20 to 100 reads). Takes one to three minutes
per human sample and needs Matrix (new in Suggests) for the largest
block of transcripts.add_uncertainty(identifiability = TRUE) runs it and adds identifiable
and identifiable_fraction to tx_uncertainty(); the table of
equivalence_identifiability() can be passed in its place.
de_confidence() then counts a non-identifiable isoform as a fourth
condition against the call (reason not identifiable from the equivalence classes) and returns an identifiable column. Off by default, so
existing results do not change.coding_potential() and
transcript_orfs() is retrained on GENCODE 50 (9,938 coding and 10,000
non-coding transcripts): held-out sensitivity 0.941 and specificity 0.919
at the 0.364 cutoff, AUC 0.978. On the same transcripts the previous
model (GENCODE 49) has 0.928, 0.936 and 0.977, so coding_probability
values change slightly and the accuracy does not.transcript_productivity() treats the GENCODE biotype
protein_coding_LoF (148 transcripts in release 50) as unproductive when
use_biotype = TRUE: a variant of the reference genome stops its
translation.write_isoformic(overwrite = TRUE) writes to a
sibling directory and swaps it in, so overwriting the directory an object
was read from no longer deletes it (and refuses directories that are not
experiments, isoformic_not_experiment_dir); de_table() and
plot_log2fc() accept results that already have gene_name, gene_id or
transcript_name columns; test_productive_fraction() and test_apa()
contrast two of several groups with levels; join_DEG_DET() uses the same
cutoff rule as the stored cutoffs (>= and <=) and gains use_fdr; a
frozen dataset that fails its checksum after download is discarded and the
next mirror is tried; plot_log2fc() no longer drops its title when a
requested feature has no data; the assay hash behind the caches is computed
once when the assays are set; tools is declared in Imports.inferential_replicates_table(), as_inferential_replicates()), and the
bundled draws are Parquet tables instead of .rds files.ESRP2 siRNA) and GSE75489 (H358 cells, ESRP1/2
shRNA), re-quantified with Salmon 2.8.0 and 30 Gibbs draws, reduced to 41
genes (under 1 MB). data-raw/knockdown/quantify_knockdown.sh and
data-raw/build_knockdown_example.R rebuild them. add_uncertainty() scales
the counts with the library_size column of col_data() when present (for
objects holding a subset of the transcripts) and accepts the table of
equivalence_support() in equivalence.read_inferential_replicates(),
equivalence_support() and salmon_run_qc() read the Gibbs/bootstrap
replicates, the equivalence classes and the run metadata;
add_uncertainty() (or read_salmon(inferential = TRUE)) stores, per
transcript, the inferential relative variance, the correlation with the
isoforms of the same gene (ambiguity groups that cannot be told apart) and
the share of uniquely mapping fragments. de_confidence() repeats a
contrast on every draw and reports the interval and sign agreement of the
fold-change with a high/moderate/low confidence class, optionally
testing ambiguous isoforms as one feature. uncertainty_report() and
uncertainty_methods() summarise the results and the settings for a
manuscript; plot_de_confidence(), plot_isoform_correlation() and
plot_inf_rv() draw them. isoform_switches(confident_only = TRUE),
dominant_isoform() and productive_fraction() use them. The tables are
written by write_isoformic(); the draws stay in memory.salmon_quant() draws 30 Gibbs samples by default (it was 100;
num_gibbs = 0 skips them).salmon_index() and salmon_quant() target Salmon 2 (the Rust rewrite;
the conda environment installs 2.8.0): options that Salmon 2 ignores
(--no-version-check, --type puff, --disableChainingHeuristic,
--useVBOpt) are no longer passed, single-end reads are passed with
--unmatedReads, and the documentation explains that indices built with
Salmon 1.x must be rebuilt.?isoformic-options documents every option and environment variable the
package reads or is affected by: the isoformic.frozen_mirrors option, the
base R download options, the cache location variables and the settings of
DuckDB and condathis.add_appris(), add_riboseq_orfs() (listed transcripts
and exon-block compatibility, riboseq_main_orf, riboseq_uorf),
add_nmd_exons(), domain_retention() (domains retained, partial or
lost relative to the canonical isoform, projected onto novel isoforms)
and gtex_reference() / gtex_tissues() (healthy-tissue productive
fraction and dominant isoform per gene).productivity_fractions() (expression share per
class and unproductive mechanism: NMD, retained intron, no ORF, non-stop),
test_productive_fraction() (quasi-binomial two-group test on counts),
productive_fraction(min_expr = , inf_reps = , reference = ) (expression
floor, uncertainty from Salmon inferential replicates, difference to the
GTEx baseline), dominant_isoform(min_expr = ) and
benchmark_productivity() (recall of NMD and retained-intron biotypes by
the rules alone, productive rate of coding and canonical transcripts).polya_sites() ranks the 3' ends of each
gene, classifies tandem_utr, alternative_last_exon and intronic
events relative to the canonical isoform, finds poly(A) signal hexamers
in the sequence and validates ends against the PolyASite atlas or the
GENCODE poly(A) GFF (download_reference(file_type = "polya"),
read_polya_gff()); apa_usage() quantifies distal usage and weighted
3' UTR length per sample or group and test_apa() tests the shift.reference_id, ref_gene_id, ref_gene_name) are kept,
and add_sqanti_classification() writes SQANTI3 structural categories
into transcript_type; tx_type_palette() includes those categories and
the unknown/novel labels. Tables are read on first use and cached keyed by file
signature, so objects sharing an annotation share one copy.write_isoformic() and
read_isoformic() (optionally self-contained with copy_annotation = TRUE).read_salmon() reads quant.sf files straight into an
IsoformicExperiment with DuckDB; read_parquet(), write_parquet() and
read_tabular() are exported; make_tx_to_gene() parses FASTA headers with
DuckDB and also accepts GFF3/GTF files and annotation directories.SummarizedExperiment-style accessors: assays(), assay(), assay<-,
assay_names(), gene_assays(), gene_assay(), gene_data(),
gene_names(), row_names(), col_names(), col_data<-, exon_table(),
annot_metadata(), clear_cache(), plus dim(), dimnames() and [.de_tx<-, de_gene<- accept DESeq2, edgeR, limma and fishpond column
names) and flagged with stored cutoffs (de_cutoffs(), de_cutoffs<-,
de_table()).isoform_fraction(),
tx_diversity() (expressed isoforms, dominant isoform, effective number of
isoforms), dominant_isoform() (per group, with dominance switches),
isoform_switches() (transcript versus gene level calls) and
exon_diversity() (constitutive versus alternative exons).as_summarized_experiment() converts back to SummarizedExperiment at the
transcript or gene level; as_isoformic() also accepts tximport lists.example_isoformic() loads a bundled experiment (preeclampsia placenta
subset, GENCODE v33) used by all examples and tests.profile_data() and plot_tx_expr() methods for the object;
plot_enrichment() and read_gmt().summarize_to_gene() is exported and sums isoforms (abundance-weighted
mean for length assays, as tximport does).vignettes/articles/data/
(not installed) and are rebuilt from the SRA reads by the scripts in
data-raw/case_studies/.inst/CITATION
carries its DOI.pixi.toml and pixi.lock define an isolated development environment
for Linux, macOS (Apple silicon) and Windows: pixi install, then
pixi run test. Linux and macOS take all R packages from conda-forge and
bioconda. On Windows, pixi run r-packages installs the packages that
conda lacks there (duckdb, condathis and the Bioconductor packages) as
binary packages from CRAN, r-universe and Bioconductor 3.22. On Linux and
macOS the environment also has LaTeX (a pinned TinyTeX release, packaged
by the recipe in .github/pixi/tinytex), so pixi run check builds the
PDF manual.add_uncertainty() reports when the replicates give fewer than 60
degrees of freedom over the study (samples x (replicates - 1); Baldoni,
Chen and Smyth 2024), the number is stored and stated by
uncertainty_methods(), and uncertainty_report() has a new replicates
table (replicate type, draws, degrees of freedom, median overdispersion).
The documentation calls the correlation rule of add_uncertainty()
"ambiguity between isoforms" instead of "identifiability", credits
mmcollapse (Turro et al. 2014) and states what the replicates do not
measure: non-identifiable abundances (Zheng et al. 2022) and estimates
that are wrong because the annotation is incomplete (Ma and Kingsford
2019).
plot_genomic_context() is a pure ggplot2 implementation reading the
exon table of the annotation (one row per transcript, introns with strand
arrows, expression labels and highlighted differentially expressed
isoforms); it no longer needs plotgardener, txdbmaker, TxDb, OrgDb
or BSgenome packages. plot_tx_context() is now an alias of its data
frame method and accepts the legacy column names.
plot_genomic_context() labels each transcript on the y axis, aligned
with its track (the labels floated above the tracks and overlapped
transcripts starting at the left edge), and prints expression labels
without padded zeros. As a plain ggplot2 object it renders the same on
every graphics device, including Cairo and the RStudio plot panel (#21).
plot_log2fc() facets several genes, orders the gene bar first and uses a
common theme; plot_log2FC() is kept as a wrapper.
plot_tx_expr() and run_enrichment() are S7 generics whose first
argument is the object (or a table); run_enrichment() no longer takes the
unused lfc_cutoff argument and expands gene sets with a single join.
prepare_profile_data() returns the same long table as profile_data()
(feature_name, feature_type, gene_name, group, mean_expr,
sd_expr, is_de); gene expression is the sum of the isoforms (it was the
mean).
prepare_annotation() and prepare_exon_annotation() read the Parquet
annotation directory and return standard column names (transcript_id,
start_pos, end_pos, ...).
Bundled example tables are now Parquet (pe_abundance.parquet,
pe_det.parquet, pe_deg.parquet, pe_col_data.parquet).
download_reference() discovers files by listing the release directory
on the server instead of relying on hard-coded file names, so any current
or future GENCODE release (and genome patch, e.g. GRCh38.p15) works.
New region argument ("all", "chr", "primary") selects the
annotation flavour; versions are normalised (49, "v49", "38" for
mouse). Errors are specific: no internet, release not found (listing the
latest releases), file not found in the release (listing the candidates),
and failed or empty transfers (partial files are removed). Defaults:
GENCODE release 50 (M39 for mouse), file_type = "gff", output_path = ":cache:".
Development tooling: jarl.toml (all jarl rules), a jarl-check GitHub
workflow, lint and test-coverage workflows, and justfile recipes aligned
with condathis.
group and levels choose the samples in the same way in every function
that has them (productive_fraction(), productivity_fractions(),
test_productive_fraction(), run_productivity_enrichment(),
apa_usage(), test_apa(), profile_data(), plot_tx_expr(),
dominant_isoform(), haplotype_summary(), test_usage(),
de_confidence(), add_uncertainty(), prepare_profile_data()): samples
of groups outside levels are ignored, as if the object had been subset
to them, so a contrast between two of several groups needs no sub-object.
The summary functions aborted in this case. Every returned value,
including as_matrix = TRUE and the inferential standard deviations,
comes from the samples kept. A level in levels without samples, or
levels without group, is an error (isoformic_group_levels); samples
without a group value are ignored with a message; factor levels without
samples are not groups. dominant_isoform() and haplotype_summary()
gain levels.
run_productivity_enrichment() gains rank_by = c("delta", "delta_z")
to rank genes by the change in productive fraction divided by its
standard deviation over the inferential replicates, and passes
include_altered, min_expr and inf_reps on to
productive_fraction(). The result has a rank_by column. delta_z is 0,
not NA, for a gene whose productive fraction changes in no replicate; a
change that is the same in every replicate is ranked with the largest
delta_z.
productive_fraction() reads the inferential replicates one at a time
and keeps only running sums, restricted to the transcripts of coding genes
and the samples of the levels, instead of holding the fractions of all
the replicates: about 18 times faster with a fraction of the memory on
295,000 transcripts, 12 samples and 100 replicates, with the same results.
inf_reps = FALSE skips the stored replicates. With the default
rank_by = "delta", run_productivity_enrichment() does not read them.
run_productivity_enrichment() also takes the table of
productive_fraction(), so that one table serves several gene set
collections.
add_uncertainty(), de_confidence(), inferential_overdispersion(),
scale_counts(), uncertainty_report() and test_usage() read the
inferential replicates one at a time from the stored list instead of
copying them into a dense array: on 534,000 transcripts, 12 samples and
100 replicates the memory above the object went from 19 GB to under 4 GB
for add_uncertainty(group = ) and de_confidence(), with the same
results.
run_enrichment() gains rank_by: besides the fold-change
("log2FC", the default) the transcripts can be ranked by
-log10(pvalue) with the sign of the fold-change ("signed_p") or by the
median fold-change over the inferential replicates divided by its standard
deviation ("lfc_z", from the confidence table of
add_uncertainty(x, group = )), so that a fold-change the quantification
cannot determine ranks low. The result has a rank_by column.
run_apa_enrichment() runs a gene set enrichment analysis on the genes
ranked by their change in distal 3' end usage or in weighted 3' UTR
length, from an object or from the table of apa_usage().
run_enrichment(), run_productivity_enrichment() and
run_apa_enrichment() take several collections of gene sets at once (a
named list of gene set lists, for example
list(hallmark = read_gmt(h), reactome = read_gmt(r))): each collection is
tested on its own and the result has a collection column, by which
plot_enrichment() splits its panels. Every gene set now needs a name: a
list with a missing or empty name is an error
(isoformic_invalid_genesets) instead of a set dropped without notice.
An enrichment without result is a table with the fgsea columns and no
rows.
productive_fraction() with inferential replicates stops when a level of
group is named sd, whose column productive_fraction_sd used to be
overwritten by the standard deviation over the replicates.
transcript_orfs() keeps the ORFs it computes in
annot_data_orfs.parquet in the annotation directory, and later calls,
sessions and objects with the same annotation read them back: for the
534,000 transcripts of GENCODE 49 the ORFs take 5 seconds instead of more
than 5 minutes, and transcript_productivity() about 35 seconds instead
of 7 minutes, once the table is written. The results are the same.
options(isoformic.orf_cache = FALSE) turns the table off.
cut_transcript_trees() reads the inferential replicates one at a time
and keeps only the mean and variance of the tree nodes, instead of a dense
array of all the replicates.
x[i, j] subsets everything stored by transcript or by sample: the
inferential replicates in metadata(x)$inf_reps and stored gene assays
follow the assays, and the uncertainty tables are restricted to the kept
transcripts. Dropping samples leaves the uncertainty tables as computed
and says so (add_uncertainty(x) recomputes them from the subset
replicates). Subscripts are checked (isoformic_invalid_subset), and must
keep at least one transcript and one sample.
test_usage() estimates the inferential overdispersion on the samples of
the test, not on all the samples of the object.
The imported packages have minimum versions, those the package is
developed and tested with: cli 3.6.6, DBI 1.3.0, dplyr 1.2.1,
duckdb 1.5.4, fs 2.1.0, ggplot2 4.0.3, ggrepel 0.9.8, glue 1.8.1,
rlang 1.3.0, S7 0.2.2, scales 1.4.0, tibble 3.3.1 and withr
3.0.3. The tests also pass with duckdb 2.0.0.
jsonlite and readr are in Suggests. The test that compares the
inferential overdispersion with edgeR::catchSalmon() needs them.
Lower memory use and faster subsetting on large objects, with the same
results. read_inferential_replicates() (and add_uncertainty(paths = ))
holds one sample at a time instead of every sample twice: its peak falls
from about 2.5 to about 1 time the size of all the replicates. x[i, j]
matches the transcripts once instead of twice per replicate (about 9
times faster with 295,000 transcripts and 20 replicates), validates the
object once, and returns the replicates without a copy when the subset
keeps everything. A subset with other transcripts gets its own cache of
derived tables, so it no longer replaces those of the object it came
from. gene_assays(), as_isoformic() (IDs without |) and
write_isoformic() no longer make spare copies of the assays.
as_summarized_experiment() leaves the inferential replicates out of the
metadata of the result.
write_isoformic(copy_annotation = TRUE) does not copy the stored ORF
table, which was not valid for the copy.
prepare_isoformic_annotation() kept only the first ##sequence-region
of a GENCODE GFF3 file (one row, chr1), because GENCODE writes one pragma
before the features of each sequence and only the header was read. The
whole file is now scanned (annot_seq_regions.parquet has 706 rows for
GENCODE 50 chr_patch_hapl_scaff), and a sequence without a pragma keeps
the largest feature end. Annotation directories made before this fix hold
the single row: rebuild them with overwrite = TRUE if you use the table.mean_inf_rv, max_inf_rv,
group_inf_rv of cut_transcript_trees()) now is the formula of
fishpond::computeInfRV(), max(var - mean, 0) / (mean + 5) + 0.01, as
the documentation said. It was max(0.01, var - mean) / (mean + 5): about
0.01 lower, and far below 0.01 for transcripts whose draws vary no more
than counts. A transcript within 0.01 below inf_rv_cutoff is now
uncertain. The tables of the uncertainty and stability articles were
rebuilt.get_isoformic_cache("sources", <name>, <version>)). With one directory
per dataset, a new version in the registry reused the cached file of the
old one, which has the same name, and rebuilt the old data.x[, j] no longer fails on an object with stored gene assays, and a
factor subscript is read as IDs, not as its integer codes.dominant_isoform() and haplotype_summary() failed with a single group.de_confidence(collapse_groups = TRUE) labelled a transcript with the ID
of its ambiguity group when it was the only member left in the object.isoform_fraction() returned cached fractions after gene_assays<-.isoform_fraction(), dominant_isoform()
and profile_data(); the matrices are now matched by sample ID.run_enrichment() and run_productivity_enrichment() run fgsea with
BiocParallel::SerialParam() unless BPPARAM or nproc is given; the
default socket cluster failed (serialize() errors) or hung on Windows
(#9).stream_fasta() drops the carriage returns of Windows (CRLF) FASTA files.build_frozen_nmd_exons() no longer replaces the Ensembl gene_id with
the StringTie GeneID of the source table.ISOFORMIC_FULL=true.ContextData, create_context_data() and prepare_annotation_db().arrow, vroom, readr, purrr, stringr, tidyr and
ggpubr; suggested packages plotgardener, txdbmaker,
GenomicFeatures, IRanges, AnnotationHub, org.Hs.eg.db,
BSgenome.Hsapiens.NCBI.GRCh38, TxDb.Hsapiens.UCSC.hg38.knownGene,
viridis, carrier and mirai.Release Date: 2025-10-06
Development Changelog: 0.1.2
parse_annotation().plot_genomic_context() function to visualize the genomic context of a
gene with its transcripts.tx_type_palette() function to provide a default color palette for
transcript types.make_tx_to_gene() output, wrong column name entrez_id replaced for tx_length.Release Date: 2024-06-18
Development Changelog: 0.1.1
download_reference(): now supports organism = c("human", "mouse"),
with organism = "human" being the default.download_reference(): argument file_type = "gtf" is the default.prepare_annotation(): Parse both GTF and GFF file formats into
required annotation data.prepare_profile_data(): accepts matrix and data.frame as input for the txi_transcript argument.Release Date: 2024-06-11
Development Changelog: 0.1.0
Release Date: 2024-06-08
Development Changelog: 0.0.1